Upload 2 files

Dockerfile

@@ -2,7 +2,8 @@ FROM python:3.10-slim
 
 ENV PYTHONUNBUFFERED=1 \
     HF_HOME=/data/.cache/huggingface \
-    OMP_NUM_THREADS=1
+    OMP_NUM_THREADS=1 \
+    PYTORCH_CUDA_ALLOC_CONF=expandable_segments:True
 
 WORKDIR /app
 

app.py

@@ -226,6 +226,15 @@ async def tts_endpoint(
         _num_samples, _sr, duration = output.get_info()
 
         audio_bytes = output.to_bytes()  # WAV bytes
+    except RuntimeError as exc:
+        # Gracefully surface CUDA OOM errors instead of crashing the app
+        message = str(exc)
+        if "CUDA out of memory" in message:
+            raise HTTPException(
+                status_code=503,
+                detail="CUDA out of memory on the Space GPU. Try shorter text, shorter speaker audio, or fewer concurrent requests.",
+            )
+        raise
     finally:
         # Cleanup temp file used for cloning (if any)
         if use_voice_cloning and speaker_path and os.path.isfile(speaker_path):